[MRG] Self-ensembling for sequence-to-sequence models #294
Conversation
Codecov Report
@@ Coverage Diff @@
## master #294 +/- ##
==========================================
+ Coverage 80.10% 80.19% +0.08%
==========================================
Files 49 49
Lines 2931 2944 +13
==========================================
+ Hits 2348 2361 +13
Misses 583 583 |
braindecode/models/util.py
Outdated
| ------- | ||
| np.ndarray : | ||
| Array of shape (n_rows, n_classes, (n_rows - 1) * stride + n_windows) | ||
| where each row is obtained by zero-padding the corresponding row in `x` |
There was a problem hiding this comment.
| where each row is obtained by zero-padding the corresponding row in `x` | |
| where each row is obtained by zero-padding the corresponding row in ``x`` |
braindecode/models/util.py
Outdated
| """ | ||
| if x.ndim != 3: | ||
| raise NotImplementedError( | ||
| f'x must be of shape (n_rows, n_clases, n_windows), got {x.shape}') |
There was a problem hiding this comment.
| f'x must be of shape (n_rows, n_clases, n_windows), got {x.shape}') | |
| f'x must be of shape (n_rows, n_classes, n_windows), got {x.shape}') |
test/unit_tests/models/test_util.py
Outdated
| n_outputs = (n_sequences - 1) * stride + n_windows | ||
| shifted_y = np.concatenate([ | ||
| np.concatenate(( | ||
| np.zeros((1, n_classes, i * stride)), dense_y[[i]], | ||
| np.zeros((1, n_classes, n_outputs - n_windows - i * stride))), | ||
| axis=2) | ||
| for i in range(n_sequences)], axis=0) | ||
| shifted_y2 = _pad_shift_array(dense_y, stride=stride) | ||
|
|
||
| assert (shifted_y == shifted_y2).all() |
There was a problem hiding this comment.
are you just testing zero arrays here? might be good to have somethign that tests actual values as well, like 1-2 more explicit tests with predefined values
There was a problem hiding this comment.
It is testing against actual values - the concatenation only adds zeros around the values in dense_y.
test/unit_tests/models/test_util.py
Outdated
| def test_aggregate_probas(n_sequences, n_classes, n_windows, stride): | ||
| n_outputs = (n_sequences - 1) * stride + n_windows | ||
| y_true = np.arange(n_outputs) % n_classes | ||
| logits = OneHotEncoder(sparse=False).fit_transform(y_true.reshape(-1, 1)) | ||
| logits = np.lib.stride_tricks.sliding_window_view( | ||
| logits, n_windows, axis=0)[::stride] | ||
|
|
||
| y_pred_probas = aggregate_probas(logits, n_windows_stride=stride) | ||
|
|
||
| assert y_pred_probas.ndim == 2 | ||
| assert y_pred_probas.shape[0] == n_outputs | ||
| assert y_pred_probas.shape[1] == n_classes | ||
| assert (y_true == y_pred_probas.argmax(axis=1)).all() |
There was a problem hiding this comment.
could you describe a bit more what you are testing here? like write a comment what is being tested? are values actually being tested? or just shapes?
There was a problem hiding this comment.
I'll add a few comments to make this clearer. But yes, this tests both shapes and values!
hubertjb
changed the title
|
maybe write a short line in whats new @hubertjb ? Then I can merge |
hubertjb
force-pushed
the
self-ensembling
branch
from
bd74581 to
8963567
Compare
|
Great, merged |
This PR introduces a utility function
aggregate_probasto perform self-ensembling as described in Phan et al. (2018). This works by aggregating the window-wise predictions of a seq2seq model when fed with overlapping sequences. This will be useful e.g. in #282.To simplify its use, I also added an attribute
file_idstoSequenceSamplerso we can easily see which file each sequence comes from.In the end, the function can be used in the following way given a seq2seq model
clf(e.g. anEEGClassifierthat wrapsUSleep):@tgnassou @agramfort